Nature Machine Intelligence'23
You will need a machine with a CUDA-enabled GPU and the Nvidia SDK installed to compile the CUDA kernels. We tested our methods on an NVIDA Tesla A100 GPU with CUDA Version 11.3.1. Moreover, the dataset requries overall 32 GB of memory.
The conda software is required for running the code. Generate a new environment with
$ conda create --name myNVSAenv python=3.7
$ conda activate myNVSAenv
We need PyTorch 1.11 and CUDA.
$ (myNVSAenv) conda install pytorch==1.11.0 torchvision==0.12.0 cudatoolkit=11.3 -c pytorch -c conda-forge
$ (myNVSAenv) pip install -r requirements.txt
As a last requirement, you need to install the neuro-vsa by cloning the code and running
$ (myNVSAenv) pip install -e /path/to/neuro-vsa/ --no-dependencies
Make sure to activate the environment before running any code. The installation time was 10 min.
Download the .npz and .xml files of the RAVEN dataset here as presented in „RAVEN: A Dataset for Relational and Analogical Visual rEasoNing„ by Zhang et al. (2019). Let DATAPATH/ be the path to the saved data.
Run the rule preprocessing script which you can get here
$ (myNVSAEnv) python3 preprocess_rule.py
You can run the end-to-end training on the RAVEN dataset for all constellation by specifying the
$ (myNVSAEnv)./run.sh
If you use the work released here for your research, please cite the preprint of this paper:
@article{hersche2022neuro,
title={A Neuro-vector-symbolic Architecture for Solving Raven's Progressive Matrices},
author={Hersche, Michael and Zeqiri, Mustafa and Benini, Luca and Sebastian, Abu and Rahimi, Abbas},
journal={Nat Mach Intell},
year={2023}
}
Please refer to the LICENSE file for the licensing of our code. Our implementation relies on PrAE released under GPL v3.0